External Validation of Dementia Risk Models in a UK Population-Based Prospective Cohort Amongst Stroke Survivors

AI Summary

Dementia risk models developed for the general population show poor discriminative performance in stroke survivors (c-statistic 0.52 to 0.69) and inconsistent calibration.
Most models offered minimal improvement over age alone as a predictor of post-stroke dementia.
Urgent development and external validation of stroke-specific dementia prediction models are required to guide early intervention and monitoring.

J Clin Epidemiol. 2026 May 21:112339. doi: 10.1016/j.jclinepi.2026.112339. Online ahead of print.

ABSTRACT

OBJECTIVES: Stroke-survivors are at a significantly elevated risk of post-stroke cognitive impairment and dementia, yet early identification of high-risk cases remains a challenge. Dementia risk models developed for the general population may not translate well to stroke populations due to unique pathophysiological mechanisms, risk factor profiles and accelerated cognitive decline often observed after cerebrovascular events. The aim of this study was to assess the accuracy of existing dementia models in a population-based stroke cohort.

STUDY DESIGN AND SETTING: This study externally validated 11 risk prediction models in a population-based cohort of stroke survivors from the EPIC-Norfolk study. Model predictive performance was assessed in line with their development parameters as well as over six follow-up periods using Cox proportional hazards and Fine and Gray competing risk models with an outcome of all-cause dementia. Model fit was evaluated for discrimination and calibration. Participants were eligible if they had experienced a stroke without any prior diagnosis of dementia.

RESULTS: In total, 3,782 stroke survivors were followed for a mean of 3.89 years (SD 4.72), with a maximum follow-up of 25.70 years, with 662 later diagnosed dementia. Model discriminative performance was generally poor across all models (c-statistic range: 0.52-0.69), with inconsistent calibration across visual and statistical assessments. Most models showed minimal predictive advantage compared to a model incorporating age as the sole predictor.

CONCLUSION: Dementia risk prediction models developed for the general population have limited transportability to stroke populations. Given the high risk of dementia post stroke there is an urgent need to develop and validate stroke-specific prediction models that account for unique risk factors in this population. This will enable clinicians to identify stroke survivors who may benefit from early intervention and enhanced monitoring to reduce their dementia risk.

PMID:42173453 | DOI:10.1016/j.jclinepi.2026.112339

Document this CPD